• I started working with Komi language in 2011
• After my MA in University of Helsinki in 2013 I worked in Germany and now in Paris
• Some work in Germany not related to my PhD
• In 2014–2016 there was a project working with one Komi dialect
• For 2017–2019 we got a continuation project

• Topic of my PhD is variation in Komi dialects
  • Relates to language contact, partly sociolinguistically motivated
  • Deliminated by …
    • … what we can say with existing data
    • … what kind of annotations we can realistically add
    • … which processes can be automatized with sufficient quality
    • … etc

• What's Komi language?
• Outcome of IKDP project
• Plans for the IKDP2 project
• Work done in Paris
• Current state of my PhD (vague deadline is in 2019)
  • Russian influence in Komi
  • Phonetic changes
  • Morphosyntactic changes
  • Better analysis of syntactic structures should follow

• Uralic language
• Three varieties: Komi-Zyrian, Komi-Permyak, Komi-Jazva
• Closely related to Udmurt, more distantly to Finnish, Hungarian etc.
• Around 300,000 speakers, everyone bilingual in Russian
• Spoken by children in several regions
• Language shift in the cities fast

• Historical expansion from south to north
• Around 10-14 varieties, each splitting to few subvarieties
• Areal spread related to the river systems
• Written language based to central Syktyvkar dialect
  • Often exhibits dialectal features
• Komi-Zyrian and Komi-Permyak mutually intelligible
• Official support of Komi-Zyrian much wider than Komi-Permyak's
• Komi-Jazva under-researched and seriously endangered, some materials exist in Russian institutions in Perm and Syktyvkar

• The project aimed to produce generally useful dataset from Iźva dialect
  • Not built with individual research question in mind
  • Tries to be areally and demographically (little bit) balanced
• Fieldwork in four locations, covering most of the areas except Siberia
• Results in almost 300,000 transcribed words
• Lots of focus in metadata and systematizing old data

• Unusually for a language documentation project, the annotations were shallow
• As far as was possible, the English or Russian translations were provided
• Rather large percentage of utterances are in Russian, or contain different Russian segments
• This is not to say that we don't need the annotations!

• The availibity of Komi materials has been growing also otherwise
• Fenno-Ugria collection has produced thousands of digitalized books and newspaper (Red Pechora etc.)
• Marina Fedina's laboratory in Syktyvkar has been proofreading and digitalizing books
• Some online corpora exist as well
• Restructuring old materials is a never ending task
  • One way to fill gaps in current data

Original layers

• Time aligned transcriptions
• Raw data, regularly adds into transcriptions

Subcorpora

• Manually corrected phoneme-aligned utterances
• Manually annotated syntactic dependencies
• Semi-manually annotated POS-tagged subcorpus

• Forced alignation to match phonemes and utterances
• Automatic segmentation tools to get materials transcribed faster
• Morphological and syntactic annotations
  • Rule-based analysators
  • Statistical SEM parser
  • Machine learning based Multilingual Bist-parser

• Majority of Iźva Komi's expansion has taken place in last 200 years
  • Komis settled to Muži in 1843, to Kola Peninsula in 1880s
• Internal variety in the dialect must be relatively new

• Usually connects to geographic isolation and/or language contact

• Phonetic differences easiest to access as the existing annotation layer captures those partially
• Within morphology there are differences in frequencies of allomorphs

• We assume there is quite much syntactic variation

• Current studies have also been test cases for the corpus -> used to select new materials

• Already studied, but there have not been corpus based investigations (Leinonen 2009)
• My recent work has discussed this from:
  • Adaptation of Russian phonology
  • Morphological changes and patterns
  • Word order changes
    • Especially written and spoken Komi differ in this
    • There are syntactic structures in Komi, which prefer OV construction although usually the word order is described as SVO
• As contact has been so long, the way borrowing takes place has strongly conventionalized

• Similar development in numerous dialects
  • Upper Sysola
  • Ińva (Komi-Permyak)
  • All have exceptionally long or intensive contact with Russian
• In Iźva (roughly) areally restricted to Siberia, Kola Peninsula, Kanin Peninsula
• Co-occurs and overlaps with other developments inside Iźva (parallels in other dialects)
  • Deaffricatization, metathesis etc.

• Uses segmentation tools developed in Institute of Phonetics and Speech Processing of the University of Munich
• Has existing models for Russian segmentation
• Can be accessed through emuR R package
• I've used it through a simple mapping between Komi and Russian phoneme inventories
  • Doesn't come without problems

• I'm currently analysing the phonetic properties of these sibilants
• The standard pronunciation on some area
• Associated with Russian

• I use the information about Russian origin of the lexical item as one variable

• The core Izhma area lacks many features that are common on tundra areas
• However, the local reindeer herders do exhibit these in their speech

• This feature is not among these ones

• Verbs in past tense can occur with several allomorphs
  • isny ~ isnys ~ iny ~ inys
  • Example: мунісны ~ мунісныс ~ муніны ~ муніныс
• To some degree areal, but variation present in many dialects
• Iźva is possibly the only dialect with all four variants in use
• Ideas to test: transitivity, presence of object, phonotaxis, syllable stress, age-graded variation, register choices, sociolinguistic differences etc.
• Current descriptions of Komi are good, but there are gaps

• Developed mainly by KyungTae Lim, based on a monolingual parser (Lim and Poibeau 2017)
• We have experimented with combining Russian and Finnish corpora into training models that are applied to Komi data
  • Results have been promising
  • For now the Finnish model has outperformed the others
• Possibly especially useful with data that contains both Russian and Komi
• Performing language identification before analysis is not very useful in this case